引導式測驗 · 硬 等級
RLHF 測驗中的分組獎勵標準化
了解 GRPO 如何比較提示組內的抽樣回應、刪除批評者以及使用 PPO 風格的目標進行策略更新。
問題 1 的 8
在分組獎勵標準化中,每個回應的獎勵與什麼進行比較?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 在分組獎勵標準化中,每個回應的獎勵與什麼進行比較?
- 其自身提示組的平均值和標準差
- 固定的全域常數
- 上一個紀元的獎勵
- 模型參數個數
- 哪種演算法與分組獎勵標準化最相關?
- 隨機梯度下降
- GRPO(群組相關策略優化)
- K-均值聚類
- 反向傳播
- GRPO 特別消除了標準 PPO 的哪些組成部分?
- 獎勵模式
- 完全神經網絡
- 獨立的價值網絡(批評者)
- 訓練資料
- 如果回應的獎勵低於其組別的平均值,會發生什麼情況?
- 它獲得積極的優勢並得到加強
- 它已從資料集中刪除
- 其獎勵加倍
- 它獲得負面優勢並被勸阻
- What does within-prompt reward normalization provide to a GRPO update?
- A relative, scaled advantage for each response compared with others for the same prompt
- A guarantee that all prompt difficulty differences disappear
- A higher learning rate
- A way to make every response receive the same reward
- What happens when every response in a prompt group receives the same reward?
- Every response gets a positive advantage
- Centering gives no relative advantage; zero spread also requires numerical handling
- The group mean doubles the reward
- The model stops needing a policy objective
- Which PPO-style update element does GRPO retain?
- A separately learned value network
- K-means clustering over tokens
- A clipped policy objective
- A language-model tokenizer
- What can a KL penalty relative to a reference policy help constrain in GRPO?
- The number of prompts in each group
- The length of the tokenizer vocabulary
- The resolution of generated images
- How far the updated policy drifts from the reference policy